1
00:00:00,000 --> 00:00:01,260
Hey everyone and welcome back.

2
00:00:01,260 --> 00:00:03,720
Today we're gonna be taking a deep dive into a paper

3
00:00:03,720 --> 00:00:07,200
that's all about making AI smarter.

4
00:00:07,200 --> 00:00:08,080
Oh, very cool.

5
00:00:08,080 --> 00:00:11,600
Yeah, it's called Mapping the Neurosymbolic AI Landscape

6
00:00:11,600 --> 00:00:12,760
by Architectures.

7
00:00:12,760 --> 00:00:13,600
Okay.

8
00:00:13,600 --> 00:00:16,520
And essentially what it's doing is trying to figure out

9
00:00:16,520 --> 00:00:20,080
how to blend the strengths of neural networks

10
00:00:20,080 --> 00:00:21,600
and symbolic logic.

11
00:00:21,600 --> 00:00:22,440
Okay, yeah.

12
00:00:22,440 --> 00:00:25,440
So kind of two different approaches to AI

13
00:00:25,440 --> 00:00:27,200
and this is kind of like a roadmap

14
00:00:27,200 --> 00:00:28,680
for how to bring them together.

15
00:00:28,680 --> 00:00:31,280
Yeah, it's really fascinating how this paper

16
00:00:31,280 --> 00:00:32,960
is addressing kind of this core challenge

17
00:00:32,960 --> 00:00:35,480
of how to combine those two things.

18
00:00:35,480 --> 00:00:36,940
The power of like neural networks

19
00:00:36,940 --> 00:00:39,480
and then the logic of symbolic reasoning.

20
00:00:39,480 --> 00:00:42,240
And it's kind of like, think of it like this,

21
00:00:42,240 --> 00:00:44,600
you know, carol networks are like these amazing artists

22
00:00:44,600 --> 00:00:47,680
that can create these super realistic paintings,

23
00:00:47,680 --> 00:00:50,080
but then they can't really explain, you know,

24
00:00:50,080 --> 00:00:52,120
how they made the choices they did to create that.

25
00:00:52,120 --> 00:00:54,200
Or like they can capture the what but not the why.

26
00:00:54,200 --> 00:00:54,960
Exactly.

27
00:00:54,960 --> 00:00:56,960
Yeah, they're good at recognizing patterns,

28
00:00:56,960 --> 00:00:59,560
but they're not so good at explaining their reasoning.

29
00:00:59,560 --> 00:01:01,460
And then on the other hand, you have symbolic logic,

30
00:01:01,460 --> 00:01:03,600
which is like a master detective, you know.

31
00:01:03,600 --> 00:01:05,160
Really good at connecting the dots

32
00:01:05,160 --> 00:01:08,520
and explaining the steps of like their deduction.

33
00:01:08,520 --> 00:01:10,720
But it can struggle when it's faced with,

34
00:01:10,720 --> 00:01:14,280
you know, massive amounts of data or complex scenarios.

35
00:01:14,280 --> 00:01:15,440
So this paper is saying,

36
00:01:15,440 --> 00:01:18,940
what if we could combine the artist's intuition

37
00:01:18,940 --> 00:01:20,280
with the detective's logic?

38
00:01:20,280 --> 00:01:21,120
Exactly.

39
00:01:21,120 --> 00:01:23,960
And that's where this idea of neurosymbolic AI comes in.

40
00:01:23,960 --> 00:01:24,960
Okay, cool.

41
00:01:24,960 --> 00:01:28,040
So the paper starts by kind of highlighting the limitations

42
00:01:28,040 --> 00:01:30,400
of just relying on neural networks alone.

43
00:01:30,400 --> 00:01:31,240
Okay.

44
00:01:31,240 --> 00:01:33,560
And one of the big ones is structured reasoning.

45
00:01:33,560 --> 00:01:37,680
So like neural networks can spot patterns and data,

46
00:01:37,680 --> 00:01:41,080
but they struggle with understanding complex relationships

47
00:01:41,080 --> 00:01:42,280
like cause and effect.

48
00:01:42,280 --> 00:01:43,400
They might see a correlation,

49
00:01:43,400 --> 00:01:45,160
but they can't really understand like

50
00:01:45,160 --> 00:01:46,600
what the underlying cause is.

51
00:01:46,600 --> 00:01:47,680
Exactly.

52
00:01:47,680 --> 00:01:50,040
Yeah, it's like noticing that people carry umbrellas

53
00:01:50,040 --> 00:01:52,640
when it's raining, but not understanding why.

54
00:01:52,640 --> 00:01:53,600
Yeah.

55
00:01:53,600 --> 00:01:56,040
Another limitation is like the huge amount

56
00:01:56,040 --> 00:01:59,240
of data neural networks need to learn effectively.

57
00:01:59,240 --> 00:02:00,080
Oh yeah.

58
00:02:00,080 --> 00:02:01,480
They're data hungry beasts.

59
00:02:01,480 --> 00:02:02,640
Yeah, for sure.

60
00:02:02,640 --> 00:02:04,880
And in a lot of real world situations,

61
00:02:04,880 --> 00:02:09,320
we just don't have tons and tons of labeled data

62
00:02:09,320 --> 00:02:10,240
to feed them.

63
00:02:10,240 --> 00:02:11,080
Right.

64
00:02:11,080 --> 00:02:12,520
It's like if you were trying to teach somebody new language

65
00:02:12,520 --> 00:02:15,160
and you just gave them a million random words,

66
00:02:15,160 --> 00:02:16,480
but with no context.

67
00:02:16,480 --> 00:02:17,320
Exactly.

68
00:02:17,320 --> 00:02:18,680
Like no grammar, no nothing.

69
00:02:18,680 --> 00:02:19,520
Yeah.

70
00:02:19,520 --> 00:02:20,680
And that brings us to another point,

71
00:02:20,680 --> 00:02:22,040
which is knowledge integration.

72
00:02:22,040 --> 00:02:22,880
Okay.

73
00:02:22,880 --> 00:02:26,800
You know, we humans come preloaded

74
00:02:26,800 --> 00:02:28,680
with tons of knowledge about the world.

75
00:02:28,680 --> 00:02:29,520
Right.

76
00:02:29,520 --> 00:02:33,400
But it's surprisingly hard to like download that knowledge

77
00:02:33,400 --> 00:02:34,520
into a neural network.

78
00:02:34,520 --> 00:02:36,800
Yeah, even if we could get enough data sometimes,

79
00:02:36,800 --> 00:02:38,480
just getting that data into the form

80
00:02:38,480 --> 00:02:40,200
that the neural network can use

81
00:02:40,200 --> 00:02:41,440
is a whole other challenge.

82
00:02:41,440 --> 00:02:42,280
Exactly.

83
00:02:42,280 --> 00:02:43,120
Yeah.

84
00:02:43,120 --> 00:02:46,320
And then the next limitation is probably the one

85
00:02:46,320 --> 00:02:47,720
that people are most familiar with,

86
00:02:47,720 --> 00:02:49,720
which is explainability.

87
00:02:49,720 --> 00:02:50,560
Okay.

88
00:02:50,560 --> 00:02:53,360
Neural networks are often seen as these black boxes.

89
00:02:53,360 --> 00:02:54,640
You know, they give us answers,

90
00:02:54,640 --> 00:02:56,880
but we don't always know how they got to those answers.

91
00:02:56,880 --> 00:02:57,720
Right.

92
00:02:57,720 --> 00:02:59,920
And especially in like healthcare or finance

93
00:02:59,920 --> 00:03:02,800
or you know, areas where you need to understand

94
00:03:02,800 --> 00:03:03,640
the reasoning.

95
00:03:03,640 --> 00:03:04,480
Exactly.

96
00:03:04,480 --> 00:03:05,760
Like you can't just trust the black box.

97
00:03:05,760 --> 00:03:06,600
Exactly.

98
00:03:06,600 --> 00:03:08,680
We need AI that we can understand and trust.

99
00:03:08,680 --> 00:03:09,520
Right.

100
00:03:09,520 --> 00:03:11,240
And then finally in neural networks,

101
00:03:11,240 --> 00:03:12,960
they don't offer guarantees.

102
00:03:12,960 --> 00:03:13,800
Okay.

103
00:03:13,800 --> 00:03:14,960
So they're probabilistic,

104
00:03:14,960 --> 00:03:17,240
meaning there's always a chance they'll make a mistake.

105
00:03:17,240 --> 00:03:18,800
So there's always a little bit of uncertainty,

106
00:03:18,800 --> 00:03:20,960
even if they're incredibly accurate.

107
00:03:20,960 --> 00:03:21,800
Right.

108
00:03:21,800 --> 00:03:23,000
And that's a problem for things like

109
00:03:23,000 --> 00:03:24,840
safety critical applications.

110
00:03:24,840 --> 00:03:25,680
Right.

111
00:03:25,680 --> 00:03:26,760
Like self-driving cars.

112
00:03:26,760 --> 00:03:27,600
Of course.

113
00:03:27,600 --> 00:03:30,480
Or you know, medical diagnosis.

114
00:03:30,480 --> 00:03:33,440
Even a small error can have huge consequences.

115
00:03:33,440 --> 00:03:35,840
So we need AI that's not just accurate,

116
00:03:35,840 --> 00:03:37,280
but also reliable.

117
00:03:37,280 --> 00:03:38,120
Exactly.

118
00:03:38,120 --> 00:03:39,640
And that's where this paper gets really interesting.

119
00:03:39,640 --> 00:03:40,480
Right.

120
00:03:40,480 --> 00:03:41,840
Because it actually lays out all these different

121
00:03:41,840 --> 00:03:44,440
neuro symbolic AI approaches.

122
00:03:44,440 --> 00:03:45,280
Right.

123
00:03:45,280 --> 00:03:48,000
That researchers are using to overcome these limitations.

124
00:03:48,000 --> 00:03:48,760
Okay.

125
00:03:48,760 --> 00:03:51,920
And these approaches generally fall into two main categories.

126
00:03:51,920 --> 00:03:54,760
Composite frameworks and monolithic frameworks.

127
00:03:54,760 --> 00:03:56,120
Okay. Let's unpack those terms.

128
00:03:56,120 --> 00:03:57,640
What are composite frameworks?

129
00:03:57,640 --> 00:03:58,720
So composite frameworks,

130
00:03:58,720 --> 00:04:01,520
you can kind of think of it as like a team effort.

131
00:04:01,520 --> 00:04:04,560
So the neural network and the symbolic logic components,

132
00:04:04,560 --> 00:04:05,600
they work together,

133
00:04:05,600 --> 00:04:07,960
but they each maintain their separate identities.

134
00:04:07,960 --> 00:04:09,920
So they're cooperating, but not merging.

135
00:04:09,920 --> 00:04:10,680
Exactly. Yeah.

136
00:04:10,680 --> 00:04:13,840
It's like two experts collaborating on a project,

137
00:04:13,840 --> 00:04:15,920
bringing their unique skills to the table.

138
00:04:15,920 --> 00:04:16,760
Yeah.

139
00:04:16,760 --> 00:04:18,000
And within composite frameworks,

140
00:04:18,000 --> 00:04:19,600
there's kind of two main types.

141
00:04:19,600 --> 00:04:20,440
Okay.

142
00:04:20,440 --> 00:04:23,040
Direct supervision and indirect supervision.

143
00:04:23,040 --> 00:04:25,920
Indirect supervision, the symbolic logic component,

144
00:04:25,920 --> 00:04:28,600
acts as a guide or coach for the neural network

145
00:04:28,600 --> 00:04:29,440
during training.

146
00:04:29,440 --> 00:04:31,440
So it's like a tutor to help the neural network learn.

147
00:04:31,440 --> 00:04:32,280
Exactly.

148
00:04:32,280 --> 00:04:33,120
Yeah.

149
00:04:33,120 --> 00:04:36,200
It can provide feedback or enforce constraints

150
00:04:36,200 --> 00:04:39,960
or even inject expert knowledge

151
00:04:39,960 --> 00:04:41,960
to help it learn better and prevent it

152
00:04:41,960 --> 00:04:44,320
from making nonsensical predictions.

153
00:04:44,320 --> 00:04:45,160
That makes sense.

154
00:04:45,160 --> 00:04:46,000
Yeah.

155
00:04:46,000 --> 00:04:47,400
So what about indirect supervision?

156
00:04:47,400 --> 00:04:48,720
And how does that work?

157
00:04:48,720 --> 00:04:50,720
So in indirect supervision,

158
00:04:50,720 --> 00:04:53,480
the neural network and symbolic logic,

159
00:04:53,480 --> 00:04:55,520
they work in a more sequential way.

160
00:04:55,520 --> 00:04:56,360
Okay.

161
00:04:56,360 --> 00:04:59,120
So the neural network focuses on pattern recognition,

162
00:04:59,120 --> 00:05:01,120
like identifying objects in an image.

163
00:05:01,120 --> 00:05:03,840
And then it passes that information to a logic model,

164
00:05:03,840 --> 00:05:06,440
which uses its reasoning abilities

165
00:05:06,440 --> 00:05:08,440
to make the final decision.

166
00:05:08,440 --> 00:05:10,120
So the neural network is like the eyes

167
00:05:10,120 --> 00:05:12,120
and the logic model is the brain.

168
00:05:12,120 --> 00:05:12,960
Exactly.

169
00:05:12,960 --> 00:05:14,000
Yeah. It's a division of labor.

170
00:05:14,000 --> 00:05:15,240
Okay. Cool.

171
00:05:15,240 --> 00:05:16,640
So within these categories,

172
00:05:16,640 --> 00:05:19,960
the paper dives into a bunch of specific frameworks.

173
00:05:19,960 --> 00:05:21,880
For example, there's one called Concordia,

174
00:05:21,880 --> 00:05:24,360
which is this really unique composite framework

175
00:05:24,360 --> 00:05:27,440
because it uses this three-way connection

176
00:05:27,440 --> 00:05:30,040
between the neural and logical components.

177
00:05:30,040 --> 00:05:31,640
Three, that sounds interesting.

178
00:05:31,640 --> 00:05:32,480
Tell me more.

179
00:05:32,480 --> 00:05:35,240
So first it allows the neural network's outputs

180
00:05:35,240 --> 00:05:37,880
to be used as priors for the logical model.

181
00:05:37,880 --> 00:05:38,720
Okay.

182
00:05:38,720 --> 00:05:39,800
So it's kind of like the neural network

183
00:05:39,800 --> 00:05:43,120
is whispering suggestions to the logic model

184
00:05:43,120 --> 00:05:45,240
based on what it sees in the data.

185
00:05:45,240 --> 00:05:46,400
So it's getting a head start.

186
00:05:46,400 --> 00:05:47,240
Exactly.

187
00:05:47,240 --> 00:05:48,800
Based on what the neural network is seeing.

188
00:05:48,800 --> 00:05:49,640
Exactly.

189
00:05:49,640 --> 00:05:52,280
And second, Concordia uses the logic model

190
00:05:52,280 --> 00:05:55,360
to inject expert knowledge into the neural network

191
00:05:55,360 --> 00:05:56,280
during training.

192
00:05:56,280 --> 00:05:58,000
Okay. So it's like having a fact checker built in.

193
00:05:58,000 --> 00:05:59,040
Exactly.

194
00:05:59,040 --> 00:06:01,600
And then third, Concordia uses something called

195
00:06:01,600 --> 00:06:04,760
a gating network to kind of dynamically combine

196
00:06:04,760 --> 00:06:06,960
the predictions of the neural and logical models

197
00:06:06,960 --> 00:06:08,320
during inference.

198
00:06:08,320 --> 00:06:10,280
Okay. So it's like a smart switchboard

199
00:06:10,280 --> 00:06:11,240
picking the best option.

200
00:06:11,240 --> 00:06:12,200
Exactly. Yeah.

201
00:06:12,200 --> 00:06:14,200
It's like choosing the most reliable output

202
00:06:14,200 --> 00:06:15,640
based on the specific input.

203
00:06:15,640 --> 00:06:16,480
That's pretty clever.

204
00:06:16,480 --> 00:06:17,320
Yeah.

205
00:06:17,320 --> 00:06:18,160
So what about the other type of framework,

206
00:06:18,160 --> 00:06:19,000
the monolithic ones?

207
00:06:19,000 --> 00:06:23,400
Right. So hashtag TTS, the deep dive episode 2024,

208
00:06:23,400 --> 00:06:25,920
1030, part two of three.

209
00:06:25,920 --> 00:06:28,680
So monolithic frameworks, they take a different approach.

210
00:06:28,680 --> 00:06:29,520
Okay.

211
00:06:29,520 --> 00:06:30,480
Instead of having these separate neural

212
00:06:30,480 --> 00:06:33,440
and symbolic components, they try to weave logic

213
00:06:33,440 --> 00:06:36,720
directly into the structure of the neural network itself.

214
00:06:36,720 --> 00:06:38,240
So it's less like a collaboration

215
00:06:38,240 --> 00:06:40,960
and more like merging their brains into one.

216
00:06:40,960 --> 00:06:41,800
Exactly.

217
00:06:41,800 --> 00:06:43,360
It's like a much tighter integration.

218
00:06:43,360 --> 00:06:44,840
Sure.

219
00:06:44,840 --> 00:06:47,040
The paper breaks these down into two types.

220
00:06:47,040 --> 00:06:47,880
Okay.

221
00:06:47,880 --> 00:06:49,120
Logically wired neural networks

222
00:06:49,120 --> 00:06:51,080
and tenserized logic programs.

223
00:06:51,080 --> 00:06:53,800
Logically wired neural networks sounds cool.

224
00:06:53,800 --> 00:06:54,640
Let's start there.

225
00:06:54,640 --> 00:06:58,040
Okay. So the idea is to create this direct mapping

226
00:06:58,040 --> 00:07:00,840
between logical statements and the neurons in the network.

227
00:07:00,840 --> 00:07:01,680
Okay.

228
00:07:01,680 --> 00:07:03,280
So if a neuron is activated, it literally means

229
00:07:03,280 --> 00:07:06,000
that a particular logical statement is true.

230
00:07:06,000 --> 00:07:08,040
And the connections between neurons

231
00:07:08,040 --> 00:07:11,080
mirror the logical relationships between those statements.

232
00:07:11,080 --> 00:07:14,360
Wow. So the network structure itself represents the logic.

233
00:07:14,360 --> 00:07:15,200
Exactly.

234
00:07:15,200 --> 00:07:16,040
That's pretty wild.

235
00:07:16,040 --> 00:07:17,600
Are there any examples of this in the paper?

236
00:07:17,600 --> 00:07:18,520
Yeah, there are.

237
00:07:18,520 --> 00:07:20,760
So one early example is K-Ban.

238
00:07:20,760 --> 00:07:21,600
Coming in.

239
00:07:21,600 --> 00:07:24,680
Which stands for Knowledge Based Artificial Neural Networks.

240
00:07:24,680 --> 00:07:25,520
Okay.

241
00:07:25,520 --> 00:07:27,120
And let's say we have a logical rule.

242
00:07:27,120 --> 00:07:30,160
Like if it's raining and you're outside,

243
00:07:30,160 --> 00:07:31,960
then you need an umbrella.

244
00:07:31,960 --> 00:07:32,800
Okay, classic.

245
00:07:32,800 --> 00:07:33,640
Right.

246
00:07:33,640 --> 00:07:35,760
So in a K-Ban, there would be neurons that represent

247
00:07:35,760 --> 00:07:38,720
each of those concepts raining outside and need umbrella.

248
00:07:38,720 --> 00:07:39,560
Okay.

249
00:07:39,560 --> 00:07:41,600
And the connections between these neurons

250
00:07:41,600 --> 00:07:43,560
would be set up to reflect that rule.

251
00:07:43,560 --> 00:07:44,300
Okay.

252
00:07:44,300 --> 00:07:46,760
So if the raining and outside neurons are activated,

253
00:07:46,760 --> 00:07:49,840
then the need umbrella neuron will also be activated.

254
00:07:49,840 --> 00:07:51,720
So you're basically like physically wiring

255
00:07:51,720 --> 00:07:52,960
the logic into the network.

256
00:07:52,960 --> 00:07:53,800
Exactly.

257
00:07:53,800 --> 00:07:55,360
And that makes it way easier to understand

258
00:07:55,360 --> 00:07:57,480
like why the network is making a decision.

259
00:07:57,480 --> 00:07:58,320
Exactly.

260
00:07:58,320 --> 00:07:59,560
Yeah, that's one of the big advantages

261
00:07:59,560 --> 00:08:01,160
of these logically wired networks

262
00:08:01,160 --> 00:08:03,040
is this inherent explainability.

263
00:08:03,040 --> 00:08:03,880
Right.

264
00:08:03,880 --> 00:08:06,120
Because you can actually see like what led to what.

265
00:08:06,120 --> 00:08:06,960
Right.

266
00:08:06,960 --> 00:08:07,800
You can trace back through the network

267
00:08:07,800 --> 00:08:11,280
and see, okay, which logical rules led to this output.

268
00:08:11,280 --> 00:08:13,560
That's a huge step forward in terms of trust

269
00:08:13,560 --> 00:08:14,640
and transparency.

270
00:08:14,640 --> 00:08:16,520
Yeah, exactly.

271
00:08:16,520 --> 00:08:18,160
But I imagine there are some limitations.

272
00:08:18,160 --> 00:08:19,120
Yeah, of course.

273
00:08:19,120 --> 00:08:20,640
So scalability is a big one.

274
00:08:20,640 --> 00:08:21,480
Okay.

275
00:08:21,480 --> 00:08:22,920
As the number of logical statements

276
00:08:22,920 --> 00:08:24,280
and relationships grows.

277
00:08:24,280 --> 00:08:25,120
Right.

278
00:08:25,120 --> 00:08:27,000
The network can become super complex

279
00:08:27,000 --> 00:08:30,520
and computationally expensive to train and use.

280
00:08:30,520 --> 00:08:34,000
So it's like trying to manage a city's traffic flow

281
00:08:34,000 --> 00:08:35,160
with a hand drawn map.

282
00:08:35,160 --> 00:08:36,000
Yeah, exactly.

283
00:08:36,000 --> 00:08:37,080
It works for small towns,

284
00:08:37,080 --> 00:08:38,920
but it gets overwhelming for big cities.

285
00:08:38,920 --> 00:08:40,360
Yeah, that's a good analogy.

286
00:08:40,360 --> 00:08:43,960
Another challenge is that real world data is often messy

287
00:08:43,960 --> 00:08:45,480
and doesn't always fit neatly

288
00:08:45,480 --> 00:08:47,840
into those predefined logical rules.

289
00:08:47,840 --> 00:08:50,680
So it's not quite as good at handling like nuance.

290
00:08:50,680 --> 00:08:51,520
Exactly.

291
00:08:51,520 --> 00:08:53,840
Okay, so what about tenserized logic programs,

292
00:08:53,840 --> 00:08:56,200
the other type of monolithic framework?

293
00:08:56,200 --> 00:08:59,080
Right, so this approach gets a little bit more mathematical.

294
00:08:59,080 --> 00:08:59,920
I'm right.

295
00:08:59,920 --> 00:09:02,640
But the core idea is to turn logical concepts

296
00:09:02,640 --> 00:09:04,240
into mathematical objects.

297
00:09:04,240 --> 00:09:05,080
Okay.

298
00:09:05,080 --> 00:09:06,240
That neural networks can work with.

299
00:09:06,240 --> 00:09:09,300
So instead of representing those logical statements

300
00:09:09,300 --> 00:09:10,840
as activated neurons,

301
00:09:10,840 --> 00:09:13,720
we represent them as matrices or tensors.

302
00:09:13,720 --> 00:09:15,360
So you're translating the language of logic

303
00:09:15,360 --> 00:09:17,080
into like the language of math.

304
00:09:17,080 --> 00:09:17,920
Exactly.

305
00:09:17,920 --> 00:09:19,160
Okay, but how does that actually work?

306
00:09:19,160 --> 00:09:20,840
So one example is tensor log.

307
00:09:20,840 --> 00:09:21,720
Tenser log.

308
00:09:21,720 --> 00:09:23,520
Which compiles logical queries

309
00:09:23,520 --> 00:09:26,000
into these things called differentiable functions.

310
00:09:26,000 --> 00:09:26,840
Okay.

311
00:09:26,840 --> 00:09:29,040
And essentially what that means is that we can use calculus

312
00:09:29,040 --> 00:09:31,000
to optimize these logical queries.

313
00:09:31,000 --> 00:09:31,840
Okay.

314
00:09:31,840 --> 00:09:34,320
Just like we use calculus to train neural networks.

315
00:09:34,320 --> 00:09:36,840
So you're kind of like blending logic and math

316
00:09:36,840 --> 00:09:40,040
in a way that lets you harness the learning power

317
00:09:40,040 --> 00:09:40,960
of neural networks.

318
00:09:40,960 --> 00:09:41,800
Exactly.

319
00:09:41,800 --> 00:09:42,920
Okay, that sounds powerful.

320
00:09:42,920 --> 00:09:44,800
Are there any downsides?

321
00:09:44,800 --> 00:09:47,120
So tensor log is known for being

322
00:09:47,120 --> 00:09:49,240
both efficient and transparent.

323
00:09:49,240 --> 00:09:52,640
So we can perform complex logical reasoning

324
00:09:52,640 --> 00:09:55,920
without sacrificing speed or our ability

325
00:09:55,920 --> 00:09:57,800
to understand how the system is working.

326
00:09:57,800 --> 00:10:00,440
So we get the best of both world speed and explainability.

327
00:10:00,440 --> 00:10:03,880
Exactly, but like all approaches, there are limitations.

328
00:10:03,880 --> 00:10:04,720
Of course.

329
00:10:04,720 --> 00:10:07,280
So tensor log works best with logical relationships

330
00:10:07,280 --> 00:10:10,720
that are kind of chain like where one conclusion leads

331
00:10:10,720 --> 00:10:12,800
to another in a very straightforward way.

332
00:10:12,800 --> 00:10:13,640
Okay.

333
00:10:13,640 --> 00:10:15,160
It might not be as effective for reasoning

334
00:10:15,160 --> 00:10:18,680
that involves really complex interconnected relationships.

335
00:10:18,680 --> 00:10:20,960
So it works great for certain types of logic,

336
00:10:20,960 --> 00:10:22,520
but not a universal solution.

337
00:10:22,520 --> 00:10:23,400
Exactly.

338
00:10:23,400 --> 00:10:24,280
Okay, cool.

339
00:10:24,280 --> 00:10:26,020
So we've explored both composite

340
00:10:26,020 --> 00:10:27,960
and monolithic frameworks.

341
00:10:27,960 --> 00:10:30,880
What are your thoughts on the overall potential

342
00:10:30,880 --> 00:10:32,880
of this neuro symbolic AI?

343
00:10:32,880 --> 00:10:35,760
Hashtag, TTS, the deep dive episode,

344
00:10:35,760 --> 00:10:39,800
2024, 1030, part three of three.

345
00:10:39,800 --> 00:10:41,560
I'm really starting to see the potential here.

346
00:10:41,560 --> 00:10:42,400
Yeah.

347
00:10:42,400 --> 00:10:43,760
It's like we're taking the best of both worlds,

348
00:10:43,760 --> 00:10:46,920
combining this like intuition and learning of neural networks

349
00:10:46,920 --> 00:10:49,440
with the reasoning of symbolic AI.

350
00:10:49,440 --> 00:10:51,400
Yeah, definitely.

351
00:10:51,400 --> 00:10:53,880
So it's not all just theory though, right?

352
00:10:53,880 --> 00:10:55,620
What are some of like the real world areas

353
00:10:55,620 --> 00:10:57,560
where you see this stuff making a difference?

354
00:10:57,560 --> 00:10:59,640
Yeah, I mean the applications are huge.

355
00:10:59,640 --> 00:11:01,000
One area is robotics.

356
00:11:01,000 --> 00:11:03,120
So you can imagine robots that can not only

357
00:11:03,120 --> 00:11:05,120
perform tasks efficiently,

358
00:11:05,120 --> 00:11:07,840
but actually understand the reasons behind those tasks.

359
00:11:07,840 --> 00:11:09,760
So they could like adapt and make decisions

360
00:11:09,760 --> 00:11:12,280
based on like a deeper understanding of what's happening.

361
00:11:12,280 --> 00:11:13,120
Exactly.

362
00:11:13,120 --> 00:11:14,160
Yeah.

363
00:11:14,160 --> 00:11:17,680
Neurosymbolic AI could enable robots to learn much faster,

364
00:11:17,680 --> 00:11:20,520
generalized to new situations better,

365
00:11:20,520 --> 00:11:23,160
and even collaborate more intuitively with humans.

366
00:11:23,160 --> 00:11:25,040
That would be huge for so many industries,

367
00:11:25,040 --> 00:11:27,680
like manufacturing healthcare space exploration.

368
00:11:27,680 --> 00:11:29,680
Oh yeah, for sure.

369
00:11:29,680 --> 00:11:30,520
What about healthcare?

370
00:11:30,520 --> 00:11:32,400
I feel like there's a lot of potential there too.

371
00:11:32,400 --> 00:11:33,240
Oh, absolutely.

372
00:11:33,240 --> 00:11:34,240
Huge potential there.

373
00:11:34,240 --> 00:11:39,240
So imagine AI systems that can analyze medical images,

374
00:11:39,640 --> 00:11:42,120
lab results, patient histories,

375
00:11:42,120 --> 00:11:46,680
to provide more accurate and explainable diagnoses.

376
00:11:46,680 --> 00:11:47,520
That would be incredible.

377
00:11:47,520 --> 00:11:48,360
And not just for the doctors,

378
00:11:48,360 --> 00:11:50,240
but also like for the patients to build trust.

379
00:11:50,240 --> 00:11:51,080
Exactly.

380
00:11:51,080 --> 00:11:51,920
With medical professionals.

381
00:11:51,920 --> 00:11:54,040
Yeah, and it goes beyond just diagnosis.

382
00:11:54,040 --> 00:11:56,440
You know, neurosymbolic AI could help

383
00:11:56,440 --> 00:11:59,020
personalize treatment plans, predict health risks,

384
00:11:59,020 --> 00:12:01,720
even assist in drug discovery.

385
00:12:01,720 --> 00:12:02,560
Wow.

386
00:12:02,560 --> 00:12:03,380
Yeah.

387
00:12:03,380 --> 00:12:04,220
What about education?

388
00:12:04,220 --> 00:12:05,400
Do you see applications there?

389
00:12:05,400 --> 00:12:06,440
Oh yeah, definitely.

390
00:12:06,440 --> 00:12:08,880
So imagine like AI tutors that can adapt

391
00:12:08,880 --> 00:12:11,520
to each student's individual learning style,

392
00:12:11,520 --> 00:12:13,440
identify their strengths and weaknesses,

393
00:12:13,440 --> 00:12:15,920
and then provide targeted feedback and support.

394
00:12:15,920 --> 00:12:17,040
That would be amazing.

395
00:12:17,040 --> 00:12:17,880
Yeah.

396
00:12:17,880 --> 00:12:19,520
They could totally revolutionize the way we learn.

397
00:12:19,520 --> 00:12:20,360
For sure.

398
00:12:20,360 --> 00:12:22,920
Make it more engaging, more effective and accessible.

399
00:12:22,920 --> 00:12:24,040
Exactly.

400
00:12:24,040 --> 00:12:26,000
And what about scientific discovery?

401
00:12:26,000 --> 00:12:28,720
Do you think this could help accelerate research?

402
00:12:28,720 --> 00:12:30,280
Oh yeah, I think it definitely could.

403
00:12:30,280 --> 00:12:33,200
So, you know, by combining that data analysis

404
00:12:33,200 --> 00:12:34,960
with logical reasoning,

405
00:12:34,960 --> 00:12:38,040
it could help scientists identify patterns,

406
00:12:38,040 --> 00:12:42,640
generate hypotheses, design experiments more effectively,

407
00:12:42,640 --> 00:12:44,320
and even lead to discoveries

408
00:12:44,320 --> 00:12:46,080
that we haven't even imagined yet.

409
00:12:46,080 --> 00:12:47,700
That's so cool.

410
00:12:47,700 --> 00:12:52,700
It really feels like we're on the edge of a new era of AI.

411
00:12:52,800 --> 00:12:55,880
One that's more powerful, more understandable,

412
00:12:55,880 --> 00:12:57,720
more aligned with human values.

413
00:12:57,720 --> 00:12:59,600
I agree. I think this paper gives us

414
00:12:59,600 --> 00:13:01,360
a really good look into that future.

415
00:13:01,360 --> 00:13:03,520
Yeah, and it highlights all the awesome progress

416
00:13:03,520 --> 00:13:04,760
that's being made in this field.

417
00:13:04,760 --> 00:13:05,600
Definitely.

418
00:13:05,600 --> 00:13:06,560
So to all our listeners out there,

419
00:13:06,560 --> 00:13:08,440
if you're interested in learning more

420
00:13:08,440 --> 00:13:10,120
about Neurosymbolic AI,

421
00:13:10,120 --> 00:13:12,520
I highly recommend checking out this paper.

422
00:13:12,520 --> 00:13:13,760
Yeah, for sure.

423
00:13:13,760 --> 00:13:15,240
And who knows, maybe you will be inspired

424
00:13:15,240 --> 00:13:16,680
to contribute to this field.

425
00:13:16,680 --> 00:13:20,160
The future of AI is full of possibilities,

426
00:13:20,160 --> 00:13:23,120
and Neurosymbolic AI is leading the way.

427
00:13:23,120 --> 00:13:24,400
Definitely.

428
00:13:24,400 --> 00:13:26,360
Well, that wraps up our deep dive

429
00:13:26,360 --> 00:13:29,880
into mapping the Neurosymbolic AI landscape.

430
00:13:29,880 --> 00:13:30,720
Thanks for joining us.

431
00:13:30,720 --> 00:13:31,560
Of course.

432
00:13:31,560 --> 00:13:32,400
And we'll see you next time.

433
00:13:32,400 --> 00:14:00,240
Later.